Sequential Discrete Hashing for Scalable Cross-Modality Similarity Retrieval

Area of Science:

• Computer Vision
• Multimedia Retrieval
• Machine Learning

Background:

• The internet's growth necessitates efficient retrieval of multimodal web data.
• Hashing methods enable fast nearest neighbor search by compressing high-dimensional data into low-dimensional Hamming spaces.
• Existing methods struggle with unified representation across different data modalities.

Purpose of the Study:

• To develop a novel supervised cross-modality hashing framework for generating unified binary codes from diverse data types.
• To improve the accuracy and efficiency of large-scale multimodal data retrieval.
• To reduce quantization errors common in hashing techniques.

Main Methods:

• Introduced Cross-Modality Sequential Discrete Hashing (CSDH), a supervised framework.
• Employed a discrete optimization scheme with a boosting strategy for sequential bit learning.
• Developed bitwise hash functions to map different modalities to unified hash codes.
• Utilized a fusion scheme for generating unified hash codes during retrieval.

Main Results:

• CSDH effectively reduces quantization errors compared to traditional rounding-off methods.
• The framework generates high-quality binary codes for multimodal data.
• Evaluated on Wiki, MIRFlickr, and NUS-WIDE datasets, CSDH demonstrated superior performance over state-of-the-art techniques.

Conclusions:

• CSDH offers a robust and effective solution for cross-modality hashing and retrieval.
• The sequential bit learning approach enhances the quality of generated hash codes.
• This method significantly advances the field of multimodal data retrieval.